Clustering the Sketch: A Novel Approach to Embedding Table Compression
This addresses memory constraints in recommendation systems, offering a novel compression method that is incremental by building on prior techniques like hashing and compositional embeddings.
The paper tackles the problem of compressing large embedding tables in recommendation systems to fit in memory during training, proposing Clustered Compositional Embeddings (CCE) that combines clustering-based compression with dynamic methods, achieving high compression rates dynamically and proving convergence with a tight iteration bound.
Embedding tables are used by machine learning systems to work with categorical features. In modern Recommendation Systems, these tables can be very large, necessitating the development of new methods for fitting them in memory, even during training. We suggest Clustered Compositional Embeddings (CCE) which combines clustering-based compression like quantization to codebooks with dynamic methods like The Hashing Trick and Compositional Embeddings (Shi et al., 2020). Experimentally CCE achieves the best of both worlds: The high compression rate of codebook-based quantization, but *dynamically* like hashing-based methods, so it can be used during training. Theoretically, we prove that CCE is guaranteed to converge to the optimal codebook and give a tight bound for the number of iterations required.